NIST: 'Safe Step' — an AI model for dynamic fire evacuation, one step at a time

NIST (the U.S. National Institute of Standards and Technology) presented on June 4, 2026 the Safe Step system, an AI model based on reinforcement learning that computes the safest fire evacuation routes in a building — one step at a time. The research was led by Hongqiang “Rory” Fang and Wai Cheong Tam.

How does Safe Step learn and decide?

Safe Step relies on reinforcement learning, an approach in which a system learns to make decisions through trial and rewards. The model learns from two sources: building floor plans and NIST’s fire simulations, which provide data on how fire and toxic gases behave in different spaces. Based on this, the system assesses which is the next safest decision when moving through a building, building the route one step at a time rather than fixing it in advance.

What does “one step at a time” mean and what is the role of sensors?

A key feature of Safe Step is its dynamism. The system uses live sensor data — temperature and air quality — from “smart” buildings equipped with sensors. It uses this data to continuously update evacuation routes and dynamic exit signs as the situation develops. Instead of a static escape map, users get instructions that change in real time, adapting to the current state of the fire and smoke.

How is Safe Step better than classic algorithms?

Traditional approaches to evacuation rely on shortest-path algorithms that seek the shortest route to an exit. Safe Step goes a step further in two ways. First, it predicts how the fire will spread, so it chooses the route not only according to the current state but also according to the expected development. Second, it optimizes for the fractional effective dose (FED) — a measure of cumulative exposure to toxic gases — choosing routes with the lowest such dose. This makes the safety of people the priority, not mere distance.

What is fractional effective dose?

Fractional effective dose (FED) is a concept from fire-safety science that expresses how much harmful gas has accumulated in a person’s body during exposure, as a fraction of the dose that would cause incapacitation. By choosing routes with the lowest FED, Safe Step seeks to minimize people’s exposure to toxic combustion products during escape. This is a fundamentally different criterion from merely avoiding the flames or seeking the nearest exit.

What are the limitations and when is application expected?

Safe Step is still in an early phase. It currently works only for single-story floor plans, while development for multi-story buildings and scenarios with multiple people at once (multi-agent) is only ongoing. NIST estimates that practical application is 5 to 10 years away. It is therefore a research achievement that shows a direction, rather than a finished system ready to be installed in buildings. Still, the combination of fire-spread prediction, live sensors, and optimization toward toxic dose offers a promising framework for future safety systems.